---
language: en
license: apache-2.0
tags:
- dora
- peft
- adapter
- finetuned
- Qwen3-32B
- medical
- clinical
- healthcare
base_model:
- Qwen/Qwen3-32B
datasets:
- TachyHealth/structured_medical
pipeline_tag: text-generation
library_name: transformers
---

# Gazal-R1-32B-sft-merged-preview

This is a DoRA adapter fine-tuned on top of [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for specialized medical reasoning tasks.

## Model description

This adapter was trained using PEFT/LoRA to enhance the base model's ability to perform step-by-step clinical reasoning and medical problem-solving. 

### Training data

The model was fine-tuned on a synthetic, structured reasoning dataset, which contains medical questions with step-by-step reasoning and final answers.

### Training procedure

The model was trained using:
- LoRA with rank 256
- DoRA (Weight-Decomposed Low-Rank Adaptation)
- rsLoRA (Rank-stabilized LoRA)
- BF16 precision training

### Use cases and limitations

This model is intended for medical education and clinical reasoning training. It should NOT be used for actual medical diagnosis or treatment decisions. Always consult qualified healthcare professionals for medical advice.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model and tokenizer
base_model_id = "Qwen/Qwen3-32B"
adapter_id = "TachyHealth/Gazal-R1-32B-sft-merged"

# Load the tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)

# Load the LoRA adapter
model = PeftModel.from_pretrained(model, adapter_id)

# Prepare a prompt following the format during training
query = """[MEDICAL QUESTION]"""

messages = [
    {"role": "system", "content": "When solving complex medical problems, follow this specific format..."},
    {"role": "user", "content": query}
]

input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

# Generate response
outputs = model.generate(
    input_ids=inputs.input_ids,
    max_new_tokens=2048,
    temperature=0.6,
    do_sample=True,
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)
```

## Performance Results

Gazal-R1 achieves exceptional performance across standard medical benchmarks:

| Model | Size | MMLU Pro (Medical) | MedMCQA | MedQA | PubMedQA |
|-------|------|-------------------|---------|-------|----------|
| [**Gazal-R1 (Final)**](https://huggingface.co/TachyHealth/Gazal-R1-32B-GRPO-preview) | **32B** | **81.6** | **71.9** | **87.1** | **79.6** |
| Gazal-R1 (SFT-only) | 32B | 79.3 | 72.3 | 86.9 | 77.6 |
| Llama 3.1 405B Instruct | 405B | 70.2 | 75.8 | 81.9 | 74.6 |
| Qwen 2.5 72B Instruct | 72B | 72.1 | 66.2 | 72.7 | 71.7 |
| Med42-Llama3.1-70B | 70B | 66.1 | 72.4 | 80.4 | 77.6 |
| Llama 3.1 70B Instruct | 70B | 74.5 | 72.5 | 78.4 | 78.5 |
| QwQ 32B | 32B | 70.1 | 65.6 | 72.3 | 73.7 |
| Qwen 3 32B | 32B | 78.4 | 71.6 | 84.4 | 76.7 |